Addressing design preferences via auto-associative connectionist models

Abstract

Truly successful designs are characterized by both satisfaction of design goals and the presence of desirable physical features. Experienced design professionals are able to exercise their cognition to satisfy both aspects to a high degree. However, complex design tasks represent challenges for human cognition, and as such computational decision support systems emerge as a relevant topic. We present a computational decision support framework for treating preferences related to physical design features. The proposed framework is based on autoassociative machine learning models that inductively learn relationships between design features characterizing highly performing designs. The knowledge matter to be learned is derived through multi-objective stochastic optimization. The resulting auto-associative models are excited with a preference vector containing a favorable composition of design features. The models are able to alleviate those relationships that result in shortcomings of performance. The model thus outputs well performing design solution, where preferences pertaining to physical features are also satisfied, to the extent possible. The paper focuses on the applicability of the proposed approach in architectural design, as an exceptional example of complex design, discusses methods to evaluate model performance, and validates the proposed method through an application focusing on the design of a sustainable façade.